B2B lead scoring:
a model that does not lie.

Every lead scoring model starts honest and ends up decorative. It begins as a way to decide who to contact first, and two quarters later it is a column in the CRM that everyone scrolls past, because the leads sitting at 85 close at roughly the same rate as the leads sitting at 55.

The usual diagnosis is that the weights need tuning. The usual reality is structural, and there are three failures stacked on top of each other: the model collapsed two different questions into one number, it was borrowed from an inbound funnel that does not exist in outbound, and nobody ever checked whether it beat random.

What a lead score is actually for

A lead score is a ranking device for a scarce resource, not a measurement of quality. Its only job is to set the order in which you spend a fixed number of first contacts. If it never changes who you touch first this week, it is reporting rather than scoring, and deleting it would cost nothing.

The scarcity is numeric. LinkedIn publishes no official ceiling on connection requests, but practitioner consensus puts the sustainable range near 100 a week, which is roughly 15 to 20 on a working day. Plan on about 100 invitations a week per account, as the usage limits page explains. The pace you actually run is what your score has to allocate, which is why separation is the only useful output. A model that puts everyone between 55 and 75 has ranked nothing at all.

Why the standard rubric is an inbound model wearing a B2B label

The published rubrics score first-party website behavior: pricing page visits, whitepaper downloads, demo requests. Those rows only exist for people who already found you. Point the same model at an outbound list and most of its scale is unavailable, so every lead compresses into a narrow band of firmographic points and nothing gets ranked.

Take a representative example. ZoomInfo's lead scoring guide, accessed August 2026, publishes a rubric of blog subscription +5, whitepaper download +25, pricing page visit +40, demo request +50, and enterprise headcount above 1,000 +30. It is a sensible model. Now ask which of those rows a cold prospect can possibly have triggered.

Four of the six rows, and the 120 points behind them, are unavailable at exactly the moment the score is supposed to do its job. What is left is a firmographic filter with a score attached, which is why so many outbound teams conclude that lead scoring "does not work for us". It is not that scoring does not work. It is that they installed a model designed to rank hands that were already raised.

The outbound version has to be built from evidence that exists before anyone has met you: public profiles, public company pages, public job posts, public engagement. That constraint is not a downgrade. It is what makes the model usable on a list you assembled yourself, and it is why the whole fit axis below can be built and tested without sending anything.

Fit and intent answer different questions

Fit asks whether an account could ever buy: headcount, industry, the problem your product removes, someone with authority over it. Intent asks whether they are in motion right now: a comment, an opened role, a job change, a funding round. Fit is durable and moves in quarters. Intent is perishable and moves in days.

That difference in speed is the whole argument. A fit score changes when the company changes. An intent score changes constantly, which is why every intent row in the rubric below carries an expiry you set yourself, and why signal-based selling treats timing as a first-class input rather than a nice-to-have.

They also fail differently. A wrong fit score sends you after companies that were never going to buy and you find out in month four. A wrong intent score sends you at the wrong moment and you find out on Tuesday. Average the two and the single number inherits both problems.

Why one blended score hides the difference

An average destroys exactly the information you need to pick an action. A lead sitting at 70 might be a perfect-fit account that is currently quiet, or a poor-fit account that is unusually loud. The first belongs on a watchlist. The second belongs nowhere. A blended score tells you to treat them identically.

There is a quieter problem too. Because intent moves daily and fit does not, a blended score wobbles for reasons unrelated to whether the account can buy. A rep watches a lead go 78, then 61, then 74 while nothing meaningful changed, and correctly concludes the number is noise. That is how scoring models die: quietly ignored rather than formally retired.

A scoring rubric you can actually run

Keep two totals out of 100 and never merge them. Fit points come from attributes you can verify without contacting anyone. Intent points come from timestamped events that expire on a schedule. Anything you cannot check against public evidence earns no points, because a signal you cannot verify is a guess wearing a number.

Four rules make that table work rather than merely exist. Cap each axis at 100, since the intent lines add to 120 on paper and a lead cannot be more than fully in motion. Give every intent line an expiry, because a score without one is a permanent record of something that stopped being true. Treat the negative fit rows as the most valuable, since removing a lead beats reordering it when the week holds 100 sends. And delete any line that has never once changed the order of your list.

Two things corrupt this table before any calibration can catch them. Proxy drift is the first: easy-to-count activity like opens and page views creeps into the fit axis and teaches it to rank the most bored people first, which is why only engagement signals with a plausible causal story earn a row. The second is that the person setting the weights is usually the person working the list, so the weights quietly become a description of who they enjoy talking to. Somebody who does not send should own the rubric.

Note what the fit half needs: public company pages, public profiles, public job posts. That is all public data, which is why the whole fit axis can be built and tested against your own list using LinkedIn's own job search for roles that imply your problem and a LinkedIn comment scraper for who is already discussing it. If the fit rows themselves are still vague, fix that first with a proper ideal customer profile, and for the wider signal catalogue see LinkedIn intent signals you can track without Sales Navigator.

Read the quadrant, not the total

Draw a line at 60 on each axis and you get four boxes. Each has exactly one correct action, and the four are not variations of one another. Two of them say do nothing this week, which is the instruction most teams skip, and skipping it is how send capacity fills permanently with the wrong people.

The top-right box should be roughly the size of your weekly send budget. That is the calibration nobody does: thresholds get picked as round numbers rather than at capacity. If 60 and 60 produces 400 people and you can send 100 invitations a week, the box is four times oversubscribed and the threshold is decorative, because a rep working alphabetically through 400 names gets the same result as a rep working the score. Raise both lines until the box fits the week. Whether that lands at 60 or at 78 is not a matter of taste, it is arithmetic on your own list.

The top-left box is what pays for the model over time, and the expected value is easy to work out. A week of 100 sends spent on generic notes returns 15 to 25 acceptances. The same 100 sends spent where a trigger gives you something specific to say returns 50 to 60. Those are the figures in LeadRiver's April 2026 review of more than 50,000 connection requests, which put trigger-based notes at 50 to 60% acceptance against 15 to 25% for generic ones, inside a typical B2B outbound band of 30 to 37%. So a watchlist entry that fires a trigger is worth two to three marginal fit-only sends. The watchlist is not a parking lot. It is the supply of your highest-value slots.

Put the idea to work

Describe who you want.
Get the people.

The people search opens one person from a name, a name and company, or a profile link, and finds many by role, company and place, in plain words. Free to try.

Search people

How many closed‑won deals you need before the model means anything

More than you have, almost certainly. Sorting wins into quintiles and eyeballing the top one proves nothing until the sample is large enough that a good result cannot be chance. With 30 closed-won deals you cannot detect anything below roughly 1.8x lift, so a genuinely useful model and a random one look identical.

The arithmetic is short enough to check yourself. Score every lead, split them into quintiles, and count how many of your wins land in the top fifth. If the model is worthless, wins scatter at random and the top quintile holds one in five of them. For n wins that null expectation is 0.2n, and the spread around it is the binomial standard deviation, the square root of n times 0.2 times 0.8, which works out to 0.4 times the square root of n. Take two standard deviations as the bar you have to clear, and the smallest lift you can distinguish from luck is that threshold divided by 0.2n.

Below about 50 wins the normal approximation is rough and those rows are indicative rather than exact, but the direction does not change with a more careful test. Three consequences follow, and none of them appears in the guides that tell you to "audit and iterate".

A model worth keeping can look like a failure. A rubric that reliably delivers 1.5x lift is worth real money on a hundred sends a week, and at 30 wins you will never see it. You will conclude scoring does not work here and quietly stop maintaining the column. That is the single commonest way a working model dies.

If you cannot reach 50 wins, calibrate on a coarser outcome. Meetings booked or opportunities created are far more frequent than closes in any pipeline, so they reach a usable sample size much sooner. You trade precision about closing for the ability to detect anything at all, which is the better trade when the alternative is a verdict you cannot trust.

Lift large enough to detect at 30 wins may not need a model. Roughly 1.8x separation is the kind of gap a single hard filter tends to produce on its own, and it is worth checking that "directors and above at companies over 200 people" does not already give you most of it before you maintain a fifteen-row rubric.

Calibrating against closed‑won

Score your closed-won deals as of the day you first contacted them, using only evidence that existed on that date, and do the same for a matched sample that never replied. If both groups land in the same fit band, the weights are decorative. Four rules keep that exercise honest.

Rebuild the score at first contact, not today. Scoring a closed-won account with everything you know now is hindsight leakage. Of course they look like a perfect fit: you learned what a perfect fit was partly from them. The only useful question is whether the model would have flagged them on day-one evidence.

Measure lift, not accuracy. Ask what share of wins landed in the top fit quintile, then check that share against the table above before you celebrate or condemn it. Top 20% of leads holding 60% of wins is 3x lift, and at 50 or more wins that is far past the bar. Top 20% holding 22% is a random number generator with a nice interface.

Judge each axis against what it claims to predict. Fit is a claim about whether a deal closes, so calibrate it on closed-won and deal size. Intent is a claim about whether a message gets answered, so calibrate it on replies inside seven days. Calibrating fit on reply rate is the commonest mistake, and it trains the model to find agreeable people rather than buyers.

Keep a random holdout. Contact only high scorers and you get outcome data only for high scorers, so the model confirms itself indefinitely. Reserve a slice of the weekly sends, ten is enough at these volumes, for mid-band leads picked at random. That slice is your only evidence about whether the threshold sits in the right place, and it is the first thing cut when a quarter gets tight.

Re-run all four quarterly. Fit weights drift as the product, the pricing and the shape of the accounts you win all change, and a model calibrated once is accurate once.

What the published lead scoring statistics actually say

Less than the pages quoting them suggest. The single most repeated figure in this category, a 192% higher lead qualification rate, comes from an Aberdeen Group survey announced in June 2008, and the same announcement reports that only 77% of the high performers it describes used scoring tools at all.

Aberdeen Group announced the survey on June 11, 2008, under the heading "Lead Scoring and Prioritization Pave the Road to Higher Conversion", and the sentence reads: "These enabling technologies help Best-in-Class companies experience a 192% higher average lead qualification rate than all other organizations." Aberdeen defined Best-in-Class as the top 20% of respondents on annual revenue growth, annual change in lead conversion rate, and annual change in cost per lead. So a group was selected on its outcomes, a second outcome was then measured on that group, and 23% of the group was not using a scoring tool. It is a cross-sectional association among 2008 survey respondents. It is not evidence that adopting a scoring model raises anyone's qualification rate, and the survey predates third-party intent data as a category and LinkedIn as an outbound channel at current scale.

The predictive scoring literature is better but narrower. The strongest recent peer-reviewed result is in Frontiers in Artificial Intelligence, March 2025, which took 23,154 CRM records from a B2B software company, cleaned them to 16,600, and compared fifteen classifiers. Gradient boosting won at 98.39% accuracy, 0.9891 AUC, 95.86% recall and 91.06% precision on a deliberately imbalanced target. That is a real finding about that dataset.

What you cannot do is carry it across, because the two strongest predictors after lead source are fields called "Reason for State" and "Lead Classification", the paper defines neither and never states when in the lead lifecycle either is populated, and it contains no discussion of leakage. Without knowing whether those values exist at scoring time or only after a rep has dispositioned the lead, the result is not reproducible on your CRM. It is good evidence that boosting beats logistic regression here. It is not evidence that a vendor's predictive score will rank your leads.

Downstream of both, the statistics roundups are worse than useless. Landbase's "30 lead scoring statistics" page, accessed August 2026, attributes all but one of its thirty figures to other marketing blogs rather than to primary research, and the exception is a 2011 Harvard Business Review article about lead response time, which is not about scoring at all.

This is why the calibration section above is not optional. There is no external number you can borrow that tells you whether your model works on your list. The only evidence that transfers is the evidence you generate.

Bottom line

One number cannot carry two questions, and an inbound rubric cannot rank a cold list. Score fit from public attributes, score intent from timestamped events that expire, act on the quadrant, set the thresholds at the capacity you actually have, and calibrate against closed-won as it looked on the day of first contact. Then hold a random slice back, so the model has something to be wrong about.

None of that lifts a reply rate by itself. It changes who receives the fixed number of messages you can send, and that is where the published variance sits: Belkins' 2026 analysis of 15.1 million touchpoints put replies at 7.9% on a cold connect-then-message sequence against 12.2% to an existing connection, and Expandi's 2026 benchmarks put LinkedIn's platform average at 10.3% against 5.1% for cold email. Belkins has since restated its cold email benchmark at 0.45% on 7.5 million 2025 sends, measured against total sends rather than opens, so treat 5.1% as the 2026-vintage figure Expandi quoted. Fit picks who, and intent picks when.

Frequently asked questions